projects / deliverable / linux.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
staging: rtl8192e: Remove pt_regs * irq handler parameter
[deliverable/linux.git] / drivers / staging / lustre / lnet / selftest / rpc.c
CommitLineData
d7e09d03
PT		 1/*
2 * GPL HEADER START
3 *
4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5 *
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
9 *
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
15 *
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
19 *
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
22 * have any questions.
23 *
24 * GPL HEADER END
25 */
26/*
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
29 *
30 * Copyright (c) 2012, Intel Corporation.
31 */
32/*
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
35 *
36 * lnet/selftest/rpc.c
37 *
38 * Author: Isaac Huang <isaac@clusterfs.com>
39 *
40 * 2012-05-13: Liang Zhen <liang@whamcloud.com>
41 * - percpt data for service to improve smp performance
42 * - code cleanup
43 */
44
45#define DEBUG_SUBSYSTEM S_LNET
46
47#include "selftest.h"
48
49typedef enum {
50 SRPC_STATE_NONE,
51 SRPC_STATE_NI_INIT,
52 SRPC_STATE_EQ_INIT,
53 SRPC_STATE_RUNNING,
54 SRPC_STATE_STOPPING,
55} srpc_state_t;
56
57struct smoketest_rpc {
58 spinlock_t rpc_glock; /* global lock */
59 srpc_service_t *rpc_services[SRPC_SERVICE_MAX_ID + 1];
60 lnet_handle_eq_t rpc_lnet_eq; /* _the_ LNet event queue */
61 srpc_state_t rpc_state;
62 srpc_counters_t rpc_counters;
63 __u64 rpc_matchbits; /* matchbits counter */
64} srpc_data;
65
66static inline int
67srpc_serv_portal(int svc_id)
68{
69 return svc_id < SRPC_FRAMEWORK_SERVICE_MAX_ID ?
70 SRPC_FRAMEWORK_REQUEST_PORTAL : SRPC_REQUEST_PORTAL;
71}
72
73/* forward ref's */
74int srpc_handle_rpc (swi_workitem_t *wi);
75
76void srpc_get_counters (srpc_counters_t *cnt)
77{
78 spin_lock(&srpc_data.rpc_glock);
79 *cnt = srpc_data.rpc_counters;
80 spin_unlock(&srpc_data.rpc_glock);
81}
82
83void srpc_set_counters (const srpc_counters_t *cnt)
84{
85 spin_lock(&srpc_data.rpc_glock);
86 srpc_data.rpc_counters = *cnt;
87 spin_unlock(&srpc_data.rpc_glock);
88}
89
90int
91srpc_add_bulk_page(srpc_bulk_t *bk, struct page *pg, int i, int nob)
92{
93 nob = min(nob, (int)PAGE_CACHE_SIZE);
94
95 LASSERT(nob > 0);
96 LASSERT(i >= 0 && i < bk->bk_niov);
97
98 bk->bk_iovs[i].kiov_offset = 0;
99 bk->bk_iovs[i].kiov_page = pg;
100 bk->bk_iovs[i].kiov_len = nob;
101 return nob;
102}
103
104void
105srpc_free_bulk (srpc_bulk_t *bk)
106{
107 int i;
108 struct page *pg;
109
110 LASSERT (bk != NULL);
111
112 for (i = 0; i < bk->bk_niov; i++) {
113 pg = bk->bk_iovs[i].kiov_page;
114 if (pg == NULL) break;
115
116 __free_page(pg);
117 }
118
119 LIBCFS_FREE(bk, offsetof(srpc_bulk_t, bk_iovs[bk->bk_niov]));
120 return;
121}
122
123srpc_bulk_t *
124srpc_alloc_bulk(int cpt, unsigned bulk_npg, unsigned bulk_len, int sink)
125{
126 srpc_bulk_t *bk;
127 struct page **pages;
128 int i;
129
130 LASSERT(bulk_npg > 0 && bulk_npg <= LNET_MAX_IOV);
131
132 LIBCFS_CPT_ALLOC(bk, lnet_cpt_table(), cpt,
133 offsetof(srpc_bulk_t, bk_iovs[bulk_npg]));
134 if (bk == NULL) {
135 CERROR("Can't allocate descriptor for %d pages\n", bulk_npg);
136 return NULL;
137 }
138
139 memset(bk, 0, offsetof(srpc_bulk_t, bk_iovs[bulk_npg]));
140 bk->bk_sink = sink;
141 bk->bk_len = bulk_len;
142 bk->bk_niov = bulk_npg;
143 UNUSED(pages);
144
145 for (i = 0; i < bulk_npg; i++) {
146 struct page *pg;
147 int nob;
148
49c02a75
PT		 149 pg = alloc_pages_node(cfs_cpt_spread_node(lnet_cpt_table(), cpt),
150 GFP_IOFS, 0);
d7e09d03
PT		 151 if (pg == NULL) {
152 CERROR("Can't allocate page %d of %d\n", i, bulk_npg);
153 srpc_free_bulk(bk);
154 return NULL;
155 }
156
157 nob = srpc_add_bulk_page(bk, pg, i, bulk_len);
158 bulk_len -= nob;
159 }
160
161 return bk;
162}
163
164static inline __u64
165srpc_next_id (void)
166{
167 __u64 id;
168
169 spin_lock(&srpc_data.rpc_glock);
170 id = srpc_data.rpc_matchbits++;
171 spin_unlock(&srpc_data.rpc_glock);
172 return id;
173}
174
175void
176srpc_init_server_rpc(struct srpc_server_rpc *rpc,
177 struct srpc_service_cd *scd,
178 struct srpc_buffer *buffer)
179{
180 memset(rpc, 0, sizeof(*rpc));
181 swi_init_workitem(&rpc->srpc_wi, rpc, srpc_handle_rpc,
182 srpc_serv_is_framework(scd->scd_svc) ?
183 lst_sched_serial : lst_sched_test[scd->scd_cpt]);
184
185 rpc->srpc_ev.ev_fired = 1; /* no event expected now */
186
187 rpc->srpc_scd = scd;
188 rpc->srpc_reqstbuf = buffer;
189 rpc->srpc_peer = buffer->buf_peer;
190 rpc->srpc_self = buffer->buf_self;
191 LNetInvalidateHandle(&rpc->srpc_replymdh);
192}
193
194static void
195srpc_service_fini(struct srpc_service *svc)
196{
197 struct srpc_service_cd *scd;
198 struct srpc_server_rpc *rpc;
199 struct srpc_buffer *buf;
200 struct list_head *q;
201 int i;
202
203 if (svc->sv_cpt_data == NULL)
204 return;
205
206 cfs_percpt_for_each(scd, i, svc->sv_cpt_data) {
207 while (1) {
208 if (!list_empty(&scd->scd_buf_posted))
209 q = &scd->scd_buf_posted;
210 else if (!list_empty(&scd->scd_buf_blocked))
211 q = &scd->scd_buf_blocked;
212 else
213 break;
214
215 while (!list_empty(q)) {
216 buf = list_entry(q->next,
217 struct srpc_buffer,
218 buf_list);
219 list_del(&buf->buf_list);
220 LIBCFS_FREE(buf, sizeof(*buf));
221 }
222 }
223
224 LASSERT(list_empty(&scd->scd_rpc_active));
225
226 while (!list_empty(&scd->scd_rpc_free)) {
227 rpc = list_entry(scd->scd_rpc_free.next,
228 struct srpc_server_rpc,
229 srpc_list);
230 list_del(&rpc->srpc_list);
231 LIBCFS_FREE(rpc, sizeof(*rpc));
232 }
233 }
234
235 cfs_percpt_free(svc->sv_cpt_data);
236 svc->sv_cpt_data = NULL;
237}
238
239static int
240srpc_service_nrpcs(struct srpc_service *svc)
241{
242 int nrpcs = svc->sv_wi_total / svc->sv_ncpts;
243
244 return srpc_serv_is_framework(svc) ?
245 max(nrpcs, SFW_FRWK_WI_MIN) : max(nrpcs, SFW_TEST_WI_MIN);
246}
247
248int srpc_add_buffer(struct swi_workitem *wi);
249
250static int
251srpc_service_init(struct srpc_service *svc)
252{
253 struct srpc_service_cd *scd;
254 struct srpc_server_rpc *rpc;
255 int nrpcs;
256 int i;
257 int j;
258
259 svc->sv_shuttingdown = 0;
260
261 svc->sv_cpt_data = cfs_percpt_alloc(lnet_cpt_table(),
262 sizeof(struct srpc_service_cd));
263 if (svc->sv_cpt_data == NULL)
264 return -ENOMEM;
265
266 svc->sv_ncpts = srpc_serv_is_framework(svc) ?
267 1 : cfs_cpt_number(lnet_cpt_table());
268 nrpcs = srpc_service_nrpcs(svc);
269
270 cfs_percpt_for_each(scd, i, svc->sv_cpt_data) {
271 scd->scd_cpt = i;
272 scd->scd_svc = svc;
273 spin_lock_init(&scd->scd_lock);
274 INIT_LIST_HEAD(&scd->scd_rpc_free);
275 INIT_LIST_HEAD(&scd->scd_rpc_active);
276 INIT_LIST_HEAD(&scd->scd_buf_posted);
277 INIT_LIST_HEAD(&scd->scd_buf_blocked);
278
279 scd->scd_ev.ev_data = scd;
280 scd->scd_ev.ev_type = SRPC_REQUEST_RCVD;
281
282 /* NB: don't use lst_sched_serial for adding buffer,
283 * see details in srpc_service_add_buffers() */
284 swi_init_workitem(&scd->scd_buf_wi, scd,
285 srpc_add_buffer, lst_sched_test[i]);
286
287 if (i != 0 && srpc_serv_is_framework(svc)) {
288 /* NB: framework service only needs srpc_service_cd for
289 * one partition, but we allocate for all to make
290 * it easier to implement, it will waste a little
291 * memory but nobody should care about this */
292 continue;
293 }
294
295 for (j = 0; j < nrpcs; j++) {
296 LIBCFS_CPT_ALLOC(rpc, lnet_cpt_table(),
297 i, sizeof(*rpc));
298 if (rpc == NULL) {
299 srpc_service_fini(svc);
300 return -ENOMEM;
301 }
302 list_add(&rpc->srpc_list, &scd->scd_rpc_free);
303 }
304 }
305
306 return 0;
307}
308
309int
310srpc_add_service(struct srpc_service *sv)
311{
312 int id = sv->sv_id;
313
314 LASSERT(0 <= id && id <= SRPC_SERVICE_MAX_ID);
315
316 if (srpc_service_init(sv) != 0)
317 return -ENOMEM;
318
319 spin_lock(&srpc_data.rpc_glock);
320
321 LASSERT(srpc_data.rpc_state == SRPC_STATE_RUNNING);
322
323 if (srpc_data.rpc_services[id] != NULL) {
324 spin_unlock(&srpc_data.rpc_glock);
325 goto failed;
326 }
327
328 srpc_data.rpc_services[id] = sv;
329 spin_unlock(&srpc_data.rpc_glock);
330
331 CDEBUG(D_NET, "Adding service: id %d, name %s\n", id, sv->sv_name);
332 return 0;
333
334 failed:
335 srpc_service_fini(sv);
336 return -EBUSY;
337}
338
339int
340srpc_remove_service (srpc_service_t *sv)
341{
342 int id = sv->sv_id;
343
344 spin_lock(&srpc_data.rpc_glock);
345
346 if (srpc_data.rpc_services[id] != sv) {
347 spin_unlock(&srpc_data.rpc_glock);
348 return -ENOENT;
349 }
350
351 srpc_data.rpc_services[id] = NULL;
352 spin_unlock(&srpc_data.rpc_glock);
353 return 0;
354}
355
356int
357srpc_post_passive_rdma(int portal, int local, __u64 matchbits, void *buf,
358 int len, int options, lnet_process_id_t peer,
359 lnet_handle_md_t *mdh, srpc_event_t *ev)
360{
361 int rc;
362 lnet_md_t md;
363 lnet_handle_me_t meh;
364
365 rc = LNetMEAttach(portal, peer, matchbits, 0, LNET_UNLINK,
366 local ? LNET_INS_LOCAL : LNET_INS_AFTER, &meh);
367 if (rc != 0) {
368 CERROR ("LNetMEAttach failed: %d\n", rc);
369 LASSERT (rc == -ENOMEM);
370 return -ENOMEM;
371 }
372
373 md.threshold = 1;
374 md.user_ptr = ev;
375 md.start = buf;
376 md.length = len;
377 md.options = options;
378 md.eq_handle = srpc_data.rpc_lnet_eq;
379
380 rc = LNetMDAttach(meh, md, LNET_UNLINK, mdh);
381 if (rc != 0) {
382 CERROR ("LNetMDAttach failed: %d\n", rc);
383 LASSERT (rc == -ENOMEM);
384
385 rc = LNetMEUnlink(meh);
386 LASSERT (rc == 0);
387 return -ENOMEM;
388 }
389
390 CDEBUG (D_NET,
391 "Posted passive RDMA: peer %s, portal %d, matchbits "LPX64"\n",
392 libcfs_id2str(peer), portal, matchbits);
393 return 0;
394}
395
396int
397srpc_post_active_rdma(int portal, __u64 matchbits, void *buf, int len,
398 int options, lnet_process_id_t peer, lnet_nid_t self,
399 lnet_handle_md_t *mdh, srpc_event_t *ev)
400{
401 int rc;
402 lnet_md_t md;
403
404 md.user_ptr = ev;
405 md.start = buf;
406 md.length = len;
407 md.eq_handle = srpc_data.rpc_lnet_eq;
408 md.threshold = ((options & LNET_MD_OP_GET) != 0) ? 2 : 1;
409 md.options = options & ~(LNET_MD_OP_PUT | LNET_MD_OP_GET);
410
411 rc = LNetMDBind(md, LNET_UNLINK, mdh);
412 if (rc != 0) {
413 CERROR ("LNetMDBind failed: %d\n", rc);
414 LASSERT (rc == -ENOMEM);
415 return -ENOMEM;
416 }
417
418 /* this is kind of an abuse of the LNET_MD_OP_{PUT,GET} options.
419 * they're only meaningful for MDs attached to an ME (i.e. passive
420 * buffers... */
421 if ((options & LNET_MD_OP_PUT) != 0) {
422 rc = LNetPut(self, *mdh, LNET_NOACK_REQ, peer,
423 portal, matchbits, 0, 0);
424 } else {
425 LASSERT ((options & LNET_MD_OP_GET) != 0);
426
427 rc = LNetGet(self, *mdh, peer, portal, matchbits, 0);
428 }
429
430 if (rc != 0) {
431 CERROR ("LNet%s(%s, %d, "LPD64") failed: %d\n",
432 ((options & LNET_MD_OP_PUT) != 0) ? "Put" : "Get",
433 libcfs_id2str(peer), portal, matchbits, rc);
434
435 /* The forthcoming unlink event will complete this operation
436 * with failure, so fall through and return success here.
437 */
438 rc = LNetMDUnlink(*mdh);
439 LASSERT (rc == 0);
440 } else {
441 CDEBUG (D_NET,
442 "Posted active RDMA: peer %s, portal %u, matchbits "LPX64"\n",
443 libcfs_id2str(peer), portal, matchbits);
444 }
445 return 0;
446}
447
448int
449srpc_post_active_rqtbuf(lnet_process_id_t peer, int service, void *buf,
450 int len, lnet_handle_md_t *mdh, srpc_event_t *ev)
451{
452 return srpc_post_active_rdma(srpc_serv_portal(service), service,
453 buf, len, LNET_MD_OP_PUT, peer,
454 LNET_NID_ANY, mdh, ev);
455}
456
457int
458srpc_post_passive_rqtbuf(int service, int local, void *buf, int len,
459 lnet_handle_md_t *mdh, srpc_event_t *ev)
460{
461 lnet_process_id_t any = {0};
462
463 any.nid = LNET_NID_ANY;
464 any.pid = LNET_PID_ANY;
465
466 return srpc_post_passive_rdma(srpc_serv_portal(service),
467 local, service, buf, len,
468 LNET_MD_OP_PUT, any, mdh, ev);
469}
470
471int
472srpc_service_post_buffer(struct srpc_service_cd *scd, struct srpc_buffer *buf)
473{
474 struct srpc_service *sv = scd->scd_svc;
475 struct srpc_msg *msg = &buf->buf_msg;
476 int rc;
477
478 LNetInvalidateHandle(&buf->buf_mdh);
479 list_add(&buf->buf_list, &scd->scd_buf_posted);
480 scd->scd_buf_nposted++;
481 spin_unlock(&scd->scd_lock);
482
483 rc = srpc_post_passive_rqtbuf(sv->sv_id,
484 !srpc_serv_is_framework(sv),
485 msg, sizeof(*msg), &buf->buf_mdh,
486 &scd->scd_ev);
487
488 /* At this point, a RPC (new or delayed) may have arrived in
489 * msg and its event handler has been called. So we must add
490 * buf to scd_buf_posted _before_ dropping scd_lock */
491
492 spin_lock(&scd->scd_lock);
493
494 if (rc == 0) {
495 if (!sv->sv_shuttingdown)
496 return 0;
497
498 spin_unlock(&scd->scd_lock);
499 /* srpc_shutdown_service might have tried to unlink me
500 * when my buf_mdh was still invalid */
501 LNetMDUnlink(buf->buf_mdh);
502 spin_lock(&scd->scd_lock);
503 return 0;
504 }
505
506 scd->scd_buf_nposted--;
507 if (sv->sv_shuttingdown)
508 return rc; /* don't allow to change scd_buf_posted */
509
510 list_del(&buf->buf_list);
511 spin_unlock(&scd->scd_lock);
512
513 LIBCFS_FREE(buf, sizeof(*buf));
514
515 spin_lock(&scd->scd_lock);
516 return rc;
517}
518
519int
520srpc_add_buffer(struct swi_workitem *wi)
521{
522 struct srpc_service_cd *scd = wi->swi_workitem.wi_data;
523 struct srpc_buffer *buf;
524 int rc = 0;
525
526 /* it's called by workitem scheduler threads, these threads
527 * should have been set CPT affinity, so buffers will be posted
528 * on CPT local list of Portal */
529 spin_lock(&scd->scd_lock);
530
531 while (scd->scd_buf_adjust > 0 &&
532 !scd->scd_svc->sv_shuttingdown) {
533 scd->scd_buf_adjust--; /* consume it */
534 scd->scd_buf_posting++;
535
536 spin_unlock(&scd->scd_lock);
537
538 LIBCFS_ALLOC(buf, sizeof(*buf));
539 if (buf == NULL) {
540 CERROR("Failed to add new buf to service: %s\n",
541 scd->scd_svc->sv_name);
542 spin_lock(&scd->scd_lock);
543 rc = -ENOMEM;
544 break;
545 }
546
547 spin_lock(&scd->scd_lock);
548 if (scd->scd_svc->sv_shuttingdown) {
549 spin_unlock(&scd->scd_lock);
550 LIBCFS_FREE(buf, sizeof(*buf));
551
552 spin_lock(&scd->scd_lock);
553 rc = -ESHUTDOWN;
554 break;
555 }
556
557 rc = srpc_service_post_buffer(scd, buf);
558 if (rc != 0)
559 break; /* buf has been freed inside */
560
561 LASSERT(scd->scd_buf_posting > 0);
562 scd->scd_buf_posting--;
563 scd->scd_buf_total++;
564 scd->scd_buf_low = MAX(2, scd->scd_buf_total / 4);
565 }
566
567 if (rc != 0) {
568 scd->scd_buf_err_stamp = cfs_time_current_sec();
569 scd->scd_buf_err = rc;
570
571 LASSERT(scd->scd_buf_posting > 0);
572 scd->scd_buf_posting--;
573 }
574
575 spin_unlock(&scd->scd_lock);
576 return 0;
577}
578
579int
580srpc_service_add_buffers(struct srpc_service *sv, int nbuffer)
581{
582 struct srpc_service_cd *scd;
583 int rc = 0;
584 int i;
585
586 LASSERTF(nbuffer > 0, "nbuffer must be positive: %d\n", nbuffer);
587
588 cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
589 spin_lock(&scd->scd_lock);
590
591 scd->scd_buf_err = 0;
592 scd->scd_buf_err_stamp = 0;
593 scd->scd_buf_posting = 0;
594 scd->scd_buf_adjust = nbuffer;
595 /* start to post buffers */
596 swi_schedule_workitem(&scd->scd_buf_wi);
597 spin_unlock(&scd->scd_lock);
598
599 /* framework service only post buffer for one partition */
600 if (srpc_serv_is_framework(sv))
601 break;
602 }
603
604 cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
605 spin_lock(&scd->scd_lock);
606 /*
607 * NB: srpc_service_add_buffers() can be called inside
608 * thread context of lst_sched_serial, and we don't normally
609 * allow to sleep inside thread context of WI scheduler
610 * because it will block current scheduler thread from doing
611 * anything else, even worse, it could deadlock if it's
612 * waiting on result from another WI of the same scheduler.
613 * However, it's safe at here because scd_buf_wi is scheduled
614 * by thread in a different WI scheduler (lst_sched_test),
615 * so we don't have any risk of deadlock, though this could
616 * block all WIs pending on lst_sched_serial for a moment
617 * which is not good but not fatal.
618 */
619 lst_wait_until(scd->scd_buf_err != 0 ||
620 (scd->scd_buf_adjust == 0 &&
621 scd->scd_buf_posting == 0),
622 scd->scd_lock, "waiting for adding buffer\n");
623
624 if (scd->scd_buf_err != 0 && rc == 0)
625 rc = scd->scd_buf_err;
626
627 spin_unlock(&scd->scd_lock);
628 }
629
630 return rc;
631}
632
633void
634srpc_service_remove_buffers(struct srpc_service *sv, int nbuffer)
635{
636 struct srpc_service_cd *scd;
637 int num;
638 int i;
639
640 LASSERT(!sv->sv_shuttingdown);
641
642 cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
643 spin_lock(&scd->scd_lock);
644
645 num = scd->scd_buf_total + scd->scd_buf_posting;
646 scd->scd_buf_adjust -= min(nbuffer, num);
647
648 spin_unlock(&scd->scd_lock);
649 }
650}
651
652/* returns 1 if sv has finished, otherwise 0 */
653int
654srpc_finish_service(struct srpc_service *sv)
655{
656 struct srpc_service_cd *scd;
657 struct srpc_server_rpc *rpc;
658 int i;
659
660 LASSERT(sv->sv_shuttingdown); /* srpc_shutdown_service called */
661
662 cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
663 spin_lock(&scd->scd_lock);
f8c47be4
SB		 664 if (!swi_deschedule_workitem(&scd->scd_buf_wi)) {
665 spin_unlock(&scd->scd_lock);
d7e09d03 666 return 0;
f8c47be4 667 }
d7e09d03
PT		 668
669 if (scd->scd_buf_nposted > 0) {
670 CDEBUG(D_NET, "waiting for %d posted buffers to unlink",
671 scd->scd_buf_nposted);
672 spin_unlock(&scd->scd_lock);
673 return 0;
674 }
675
676 if (list_empty(&scd->scd_rpc_active)) {
677 spin_unlock(&scd->scd_lock);
678 continue;
679 }
680
681 rpc = list_entry(scd->scd_rpc_active.next,
682 struct srpc_server_rpc, srpc_list);
683 CNETERR("Active RPC %p on shutdown: sv %s, peer %s, "
684 "wi %s scheduled %d running %d, "
685 "ev fired %d type %d status %d lnet %d\n",
686 rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
687 swi_state2str(rpc->srpc_wi.swi_state),
688 rpc->srpc_wi.swi_workitem.wi_scheduled,
689 rpc->srpc_wi.swi_workitem.wi_running,
690 rpc->srpc_ev.ev_fired, rpc->srpc_ev.ev_type,
691 rpc->srpc_ev.ev_status, rpc->srpc_ev.ev_lnet);
692 spin_unlock(&scd->scd_lock);
693 return 0;
694 }
695
696 /* no lock needed from now on */
697 srpc_service_fini(sv);
698 return 1;
699}
700
701/* called with sv->sv_lock held */
702void
703srpc_service_recycle_buffer(struct srpc_service_cd *scd, srpc_buffer_t *buf)
704{
705 if (!scd->scd_svc->sv_shuttingdown && scd->scd_buf_adjust >= 0) {
706 if (srpc_service_post_buffer(scd, buf) != 0) {
707 CWARN("Failed to post %s buffer\n",
708 scd->scd_svc->sv_name);
709 }
710 return;
711 }
712
713 /* service is shutting down, or we want to recycle some buffers */
714 scd->scd_buf_total--;
715
716 if (scd->scd_buf_adjust < 0) {
717 scd->scd_buf_adjust++;
718 if (scd->scd_buf_adjust < 0 &&
719 scd->scd_buf_total == 0 && scd->scd_buf_posting == 0) {
720 CDEBUG(D_INFO,
721 "Try to recyle %d buffers but nothing left\n",
722 scd->scd_buf_adjust);
723 scd->scd_buf_adjust = 0;
724 }
725 }
726
727 spin_unlock(&scd->scd_lock);
728 LIBCFS_FREE(buf, sizeof(*buf));
729 spin_lock(&scd->scd_lock);
730}
731
732void
733srpc_abort_service(struct srpc_service *sv)
734{
735 struct srpc_service_cd *scd;
736 struct srpc_server_rpc *rpc;
737 int i;
738
739 CDEBUG(D_NET, "Aborting service: id %d, name %s\n",
740 sv->sv_id, sv->sv_name);
741
742 cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
743 spin_lock(&scd->scd_lock);
744
745 /* schedule in-flight RPCs to notice the abort, NB:
746 * racing with incoming RPCs; complete fix should make test
747 * RPCs carry session ID in its headers */
748 list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list) {
749 rpc->srpc_aborted = 1;
750 swi_schedule_workitem(&rpc->srpc_wi);
751 }
752
753 spin_unlock(&scd->scd_lock);
754 }
755}
756
757void
758srpc_shutdown_service(srpc_service_t *sv)
759{
760 struct srpc_service_cd *scd;
761 struct srpc_server_rpc *rpc;
762 srpc_buffer_t *buf;
763 int i;
764
765 CDEBUG(D_NET, "Shutting down service: id %d, name %s\n",
766 sv->sv_id, sv->sv_name);
767
768 cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
769 spin_lock(&scd->scd_lock);
770
771 sv->sv_shuttingdown = 1; /* i.e. no new active RPC */
772
773 cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
774 spin_unlock(&scd->scd_lock);
775
776 cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
777 spin_lock(&scd->scd_lock);
778
779 /* schedule in-flight RPCs to notice the shutdown */
780 list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list)
781 swi_schedule_workitem(&rpc->srpc_wi);
782
783 spin_unlock(&scd->scd_lock);
784
785 /* OK to traverse scd_buf_posted without lock, since no one
786 * touches scd_buf_posted now */
787 list_for_each_entry(buf, &scd->scd_buf_posted, buf_list)
788 LNetMDUnlink(buf->buf_mdh);
789 }
790}
791
792int
793srpc_send_request (srpc_client_rpc_t *rpc)
794{
795 srpc_event_t *ev = &rpc->crpc_reqstev;
796 int rc;
797
798 ev->ev_fired = 0;
799 ev->ev_data = rpc;
800 ev->ev_type = SRPC_REQUEST_SENT;
801
802 rc = srpc_post_active_rqtbuf(rpc->crpc_dest, rpc->crpc_service,
803 &rpc->crpc_reqstmsg, sizeof(srpc_msg_t),
804 &rpc->crpc_reqstmdh, ev);
805 if (rc != 0) {
806 LASSERT (rc == -ENOMEM);
807 ev->ev_fired = 1; /* no more event expected */
808 }
809 return rc;
810}
811
812int
813srpc_prepare_reply (srpc_client_rpc_t *rpc)
814{
815 srpc_event_t *ev = &rpc->crpc_replyev;
816 __u64 *id = &rpc->crpc_reqstmsg.msg_body.reqst.rpyid;
817 int rc;
818
819 ev->ev_fired = 0;
820 ev->ev_data = rpc;
821 ev->ev_type = SRPC_REPLY_RCVD;
822
823 *id = srpc_next_id();
824
825 rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
826 &rpc->crpc_replymsg, sizeof(srpc_msg_t),
827 LNET_MD_OP_PUT, rpc->crpc_dest,
828 &rpc->crpc_replymdh, ev);
829 if (rc != 0) {
830 LASSERT (rc == -ENOMEM);
831 ev->ev_fired = 1; /* no more event expected */
832 }
833 return rc;
834}
835
836int
837srpc_prepare_bulk (srpc_client_rpc_t *rpc)
838{
839 srpc_bulk_t *bk = &rpc->crpc_bulk;
840 srpc_event_t *ev = &rpc->crpc_bulkev;
841 __u64 *id = &rpc->crpc_reqstmsg.msg_body.reqst.bulkid;
842 int rc;
843 int opt;
844
845 LASSERT (bk->bk_niov <= LNET_MAX_IOV);
846
847 if (bk->bk_niov == 0) return 0; /* nothing to do */
848
849 opt = bk->bk_sink ? LNET_MD_OP_PUT : LNET_MD_OP_GET;
850 opt |= LNET_MD_KIOV;
851
852 ev->ev_fired = 0;
853 ev->ev_data = rpc;
854 ev->ev_type = SRPC_BULK_REQ_RCVD;
855
856 *id = srpc_next_id();
857
858 rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
859 &bk->bk_iovs[0], bk->bk_niov, opt,
860 rpc->crpc_dest, &bk->bk_mdh, ev);
861 if (rc != 0) {
862 LASSERT (rc == -ENOMEM);
863 ev->ev_fired = 1; /* no more event expected */
864 }
865 return rc;
866}
867
868int
869srpc_do_bulk (srpc_server_rpc_t *rpc)
870{
871 srpc_event_t *ev = &rpc->srpc_ev;
872 srpc_bulk_t *bk = rpc->srpc_bulk;
873 __u64 id = rpc->srpc_reqstbuf->buf_msg.msg_body.reqst.bulkid;
874 int rc;
875 int opt;
876
877 LASSERT (bk != NULL);
878
879 opt = bk->bk_sink ? LNET_MD_OP_GET : LNET_MD_OP_PUT;
880 opt |= LNET_MD_KIOV;
881
882 ev->ev_fired = 0;
883 ev->ev_data = rpc;
884 ev->ev_type = bk->bk_sink ? SRPC_BULK_GET_RPLD : SRPC_BULK_PUT_SENT;
885
886 rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, id,
887 &bk->bk_iovs[0], bk->bk_niov, opt,
888 rpc->srpc_peer, rpc->srpc_self,
889 &bk->bk_mdh, ev);
890 if (rc != 0)
891 ev->ev_fired = 1; /* no more event expected */
892 return rc;
893}
894
895/* only called from srpc_handle_rpc */
896void
897srpc_server_rpc_done(srpc_server_rpc_t *rpc, int status)
898{
899 struct srpc_service_cd *scd = rpc->srpc_scd;
900 struct srpc_service *sv = scd->scd_svc;
901 srpc_buffer_t *buffer;
902
903 LASSERT (status != 0 || rpc->srpc_wi.swi_state == SWI_STATE_DONE);
904
905 rpc->srpc_status = status;
906
907 CDEBUG_LIMIT (status == 0 ? D_NET : D_NETERROR,
908 "Server RPC %p done: service %s, peer %s, status %s:%d\n",
909 rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
910 swi_state2str(rpc->srpc_wi.swi_state), status);
911
912 if (status != 0) {
913 spin_lock(&srpc_data.rpc_glock);
914 srpc_data.rpc_counters.rpcs_dropped++;
915 spin_unlock(&srpc_data.rpc_glock);
916 }
917
918 if (rpc->srpc_done != NULL)
919 (*rpc->srpc_done) (rpc);
920 LASSERT(rpc->srpc_bulk == NULL);
921
922 spin_lock(&scd->scd_lock);
923
924 if (rpc->srpc_reqstbuf != NULL) {
925 /* NB might drop sv_lock in srpc_service_recycle_buffer, but
926 * sv won't go away for scd_rpc_active must not be empty */
927 srpc_service_recycle_buffer(scd, rpc->srpc_reqstbuf);
928 rpc->srpc_reqstbuf = NULL;
929 }
930
931 list_del(&rpc->srpc_list); /* from scd->scd_rpc_active */
932
933 /*
934 * No one can schedule me now since:
935 * - I'm not on scd_rpc_active.
936 * - all LNet events have been fired.
937 * Cancel pending schedules and prevent future schedule attempts:
938 */
939 LASSERT(rpc->srpc_ev.ev_fired);
940 swi_exit_workitem(&rpc->srpc_wi);
941
942 if (!sv->sv_shuttingdown && !list_empty(&scd->scd_buf_blocked)) {
943 buffer = list_entry(scd->scd_buf_blocked.next,
944 srpc_buffer_t, buf_list);
945 list_del(&buffer->buf_list);
946
947 srpc_init_server_rpc(rpc, scd, buffer);
948 list_add_tail(&rpc->srpc_list, &scd->scd_rpc_active);
949 swi_schedule_workitem(&rpc->srpc_wi);
950 } else {
951 list_add(&rpc->srpc_list, &scd->scd_rpc_free);
952 }
953
954 spin_unlock(&scd->scd_lock);
955 return;
956}
957
958/* handles an incoming RPC */
959int
960srpc_handle_rpc(swi_workitem_t *wi)
961{
962 struct srpc_server_rpc *rpc = wi->swi_workitem.wi_data;
963 struct srpc_service_cd *scd = rpc->srpc_scd;
964 struct srpc_service *sv = scd->scd_svc;
965 srpc_event_t *ev = &rpc->srpc_ev;
966 int rc = 0;
967
968 LASSERT(wi == &rpc->srpc_wi);
969
970 spin_lock(&scd->scd_lock);
971
972 if (sv->sv_shuttingdown || rpc->srpc_aborted) {
973 spin_unlock(&scd->scd_lock);
974
975 if (rpc->srpc_bulk != NULL)
976 LNetMDUnlink(rpc->srpc_bulk->bk_mdh);
977 LNetMDUnlink(rpc->srpc_replymdh);
978
979 if (ev->ev_fired) { /* no more event, OK to finish */
980 srpc_server_rpc_done(rpc, -ESHUTDOWN);
981 return 1;
982 }
983 return 0;
984 }
985
986 spin_unlock(&scd->scd_lock);
987
988 switch (wi->swi_state) {
989 default:
990 LBUG ();
991 case SWI_STATE_NEWBORN: {
992 srpc_msg_t *msg;
993 srpc_generic_reply_t *reply;
994
995 msg = &rpc->srpc_reqstbuf->buf_msg;
996 reply = &rpc->srpc_replymsg.msg_body.reply;
997
998 if (msg->msg_magic == 0) {
999 /* moaned already in srpc_lnet_ev_handler */
1000 srpc_server_rpc_done(rpc, EBADMSG);
1001 return 1;
1002 }
1003
1004 srpc_unpack_msg_hdr(msg);
1005 if (msg->msg_version != SRPC_MSG_VERSION) {
1006 CWARN("Version mismatch: %u, %u expected, from %s\n",
1007 msg->msg_version, SRPC_MSG_VERSION,
1008 libcfs_id2str(rpc->srpc_peer));
1009 reply->status = EPROTO;
1010 /* drop through and send reply */
1011 } else {
1012 reply->status = 0;
1013 rc = (*sv->sv_handler)(rpc);
1014 LASSERT(reply->status == 0 || !rpc->srpc_bulk);
1015 if (rc != 0) {
1016 srpc_server_rpc_done(rpc, rc);
1017 return 1;
1018 }
1019 }
1020
1021 wi->swi_state = SWI_STATE_BULK_STARTED;
1022
1023 if (rpc->srpc_bulk != NULL) {
1024 rc = srpc_do_bulk(rpc);
1025 if (rc == 0)
1026 return 0; /* wait for bulk */
1027
1028 LASSERT (ev->ev_fired);
1029 ev->ev_status = rc;
1030 }
1031 }
1032 case SWI_STATE_BULK_STARTED:
1033 LASSERT (rpc->srpc_bulk == NULL || ev->ev_fired);
1034
1035 if (rpc->srpc_bulk != NULL) {
1036 rc = ev->ev_status;
1037
1038 if (sv->sv_bulk_ready != NULL)
1039 rc = (*sv->sv_bulk_ready) (rpc, rc);
1040
1041 if (rc != 0) {
1042 srpc_server_rpc_done(rpc, rc);
1043 return 1;
1044 }
1045 }
1046
1047 wi->swi_state = SWI_STATE_REPLY_SUBMITTED;
1048 rc = srpc_send_reply(rpc);
1049 if (rc == 0)
1050 return 0; /* wait for reply */
1051 srpc_server_rpc_done(rpc, rc);
1052 return 1;
1053
1054 case SWI_STATE_REPLY_SUBMITTED:
1055 if (!ev->ev_fired) {
1056 CERROR("RPC %p: bulk %p, service %d\n",
1057 rpc, rpc->srpc_bulk, sv->sv_id);
1058 CERROR("Event: status %d, type %d, lnet %d\n",
1059 ev->ev_status, ev->ev_type, ev->ev_lnet);
1060 LASSERT (ev->ev_fired);
1061 }
1062
1063 wi->swi_state = SWI_STATE_DONE;
1064 srpc_server_rpc_done(rpc, ev->ev_status);
1065 return 1;
1066 }
1067
1068 return 0;
1069}
1070
1071void
1072srpc_client_rpc_expired (void *data)
1073{
1074 srpc_client_rpc_t *rpc = data;
1075
1076 CWARN ("Client RPC expired: service %d, peer %s, timeout %d.\n",
1077 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1078 rpc->crpc_timeout);
1079
1080 spin_lock(&rpc->crpc_lock);
1081
1082 rpc->crpc_timeout = 0;
1083 srpc_abort_rpc(rpc, -ETIMEDOUT);
1084
1085 spin_unlock(&rpc->crpc_lock);
1086
1087 spin_lock(&srpc_data.rpc_glock);
1088 srpc_data.rpc_counters.rpcs_expired++;
1089 spin_unlock(&srpc_data.rpc_glock);
1090}
1091
1092inline void
1093srpc_add_client_rpc_timer (srpc_client_rpc_t *rpc)
1094{
1095 stt_timer_t *timer = &rpc->crpc_timer;
1096
1097 if (rpc->crpc_timeout == 0) return;
1098
1099 INIT_LIST_HEAD(&timer->stt_list);
1100 timer->stt_data = rpc;
1101 timer->stt_func = srpc_client_rpc_expired;
1102 timer->stt_expires = cfs_time_add(rpc->crpc_timeout,
1103 cfs_time_current_sec());
1104 stt_add_timer(timer);
1105 return;
1106}
1107
1108/*
1109 * Called with rpc->crpc_lock held.
1110 *
1111 * Upon exit the RPC expiry timer is not queued and the handler is not
1112 * running on any CPU. */
1113void
1114srpc_del_client_rpc_timer (srpc_client_rpc_t *rpc)
1115{
1116 /* timer not planted or already exploded */
1117 if (rpc->crpc_timeout == 0)
1118 return;
1119
1120 /* timer sucessfully defused */
1121 if (stt_del_timer(&rpc->crpc_timer))
1122 return;
1123
1124 /* timer detonated, wait for it to explode */
1125 while (rpc->crpc_timeout != 0) {
1126 spin_unlock(&rpc->crpc_lock);
1127
1128 schedule();
1129
1130 spin_lock(&rpc->crpc_lock);
1131 }
1132}
1133
1134void
1135srpc_client_rpc_done (srpc_client_rpc_t *rpc, int status)
1136{
1137 swi_workitem_t *wi = &rpc->crpc_wi;
1138
1139 LASSERT(status != 0 || wi->swi_state == SWI_STATE_DONE);
1140
1141 spin_lock(&rpc->crpc_lock);
1142
1143 rpc->crpc_closed = 1;
1144 if (rpc->crpc_status == 0)
1145 rpc->crpc_status = status;
1146
1147 srpc_del_client_rpc_timer(rpc);
1148
1149 CDEBUG_LIMIT ((status == 0) ? D_NET : D_NETERROR,
1150 "Client RPC done: service %d, peer %s, status %s:%d:%d\n",
1151 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1152 swi_state2str(wi->swi_state), rpc->crpc_aborted, status);
1153
1154 /*
1155 * No one can schedule me now since:
1156 * - RPC timer has been defused.
1157 * - all LNet events have been fired.
1158 * - crpc_closed has been set, preventing srpc_abort_rpc from
1159 * scheduling me.
1160 * Cancel pending schedules and prevent future schedule attempts:
1161 */
1162 LASSERT (!srpc_event_pending(rpc));
1163 swi_exit_workitem(wi);
1164
1165 spin_unlock(&rpc->crpc_lock);
1166
1167 (*rpc->crpc_done)(rpc);
1168 return;
1169}
1170
1171/* sends an outgoing RPC */
1172int
1173srpc_send_rpc (swi_workitem_t *wi)
1174{
1175 int rc = 0;
1176 srpc_client_rpc_t *rpc;
1177 srpc_msg_t *reply;
1178 int do_bulk;
1179
1180 LASSERT(wi != NULL);
1181
1182 rpc = wi->swi_workitem.wi_data;
1183
1184 LASSERT (rpc != NULL);
1185 LASSERT (wi == &rpc->crpc_wi);
1186
1187 reply = &rpc->crpc_replymsg;
1188 do_bulk = rpc->crpc_bulk.bk_niov > 0;
1189
1190 spin_lock(&rpc->crpc_lock);
1191
1192 if (rpc->crpc_aborted) {
1193 spin_unlock(&rpc->crpc_lock);
1194 goto abort;
1195 }
1196
1197 spin_unlock(&rpc->crpc_lock);
1198
1199 switch (wi->swi_state) {
1200 default:
1201 LBUG ();
1202 case SWI_STATE_NEWBORN:
1203 LASSERT (!srpc_event_pending(rpc));
1204
1205 rc = srpc_prepare_reply(rpc);
1206 if (rc != 0) {
1207 srpc_client_rpc_done(rpc, rc);
1208 return 1;
1209 }
1210
1211 rc = srpc_prepare_bulk(rpc);
1212 if (rc != 0) break;
1213
1214 wi->swi_state = SWI_STATE_REQUEST_SUBMITTED;
1215 rc = srpc_send_request(rpc);
1216 break;
1217
1218 case SWI_STATE_REQUEST_SUBMITTED:
1219 /* CAVEAT EMPTOR: rqtev, rpyev, and bulkev may come in any
1220 * order; however, they're processed in a strict order:
1221 * rqt, rpy, and bulk. */
1222 if (!rpc->crpc_reqstev.ev_fired) break;
1223
1224 rc = rpc->crpc_reqstev.ev_status;
1225 if (rc != 0) break;
1226
1227 wi->swi_state = SWI_STATE_REQUEST_SENT;
1228 /* perhaps more events, fall thru */
1229 case SWI_STATE_REQUEST_SENT: {
1230 srpc_msg_type_t type = srpc_service2reply(rpc->crpc_service);
1231
1232 if (!rpc->crpc_replyev.ev_fired) break;
1233
1234 rc = rpc->crpc_replyev.ev_status;
1235 if (rc != 0) break;
1236
1237 srpc_unpack_msg_hdr(reply);
1238 if (reply->msg_type != type ||
1239 (reply->msg_magic != SRPC_MSG_MAGIC &&
1240 reply->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1241 CWARN ("Bad message from %s: type %u (%d expected),"
1242 " magic %u (%d expected).\n",
1243 libcfs_id2str(rpc->crpc_dest),
1244 reply->msg_type, type,
1245 reply->msg_magic, SRPC_MSG_MAGIC);
1246 rc = -EBADMSG;
1247 break;
1248 }
1249
1250 if (do_bulk && reply->msg_body.reply.status != 0) {
1251 CWARN ("Remote error %d at %s, unlink bulk buffer in "
1252 "case peer didn't initiate bulk transfer\n",
1253 reply->msg_body.reply.status,
1254 libcfs_id2str(rpc->crpc_dest));
1255 LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1256 }
1257
1258 wi->swi_state = SWI_STATE_REPLY_RECEIVED;
1259 }
1260 case SWI_STATE_REPLY_RECEIVED:
1261 if (do_bulk && !rpc->crpc_bulkev.ev_fired) break;
1262
1263 rc = do_bulk ? rpc->crpc_bulkev.ev_status : 0;
1264
1265 /* Bulk buffer was unlinked due to remote error. Clear error
1266 * since reply buffer still contains valid data.
1267 * NB rpc->crpc_done shouldn't look into bulk data in case of
1268 * remote error. */
1269 if (do_bulk && rpc->crpc_bulkev.ev_lnet == LNET_EVENT_UNLINK &&
1270 rpc->crpc_status == 0 && reply->msg_body.reply.status != 0)
1271 rc = 0;
1272
1273 wi->swi_state = SWI_STATE_DONE;
1274 srpc_client_rpc_done(rpc, rc);
1275 return 1;
1276 }
1277
1278 if (rc != 0) {
1279 spin_lock(&rpc->crpc_lock);
1280 srpc_abort_rpc(rpc, rc);
1281 spin_unlock(&rpc->crpc_lock);
1282 }
1283
1284abort:
1285 if (rpc->crpc_aborted) {
1286 LNetMDUnlink(rpc->crpc_reqstmdh);
1287 LNetMDUnlink(rpc->crpc_replymdh);
1288 LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1289
1290 if (!srpc_event_pending(rpc)) {
1291 srpc_client_rpc_done(rpc, -EINTR);
1292 return 1;
1293 }
1294 }
1295 return 0;
1296}
1297
1298srpc_client_rpc_t *
1299srpc_create_client_rpc (lnet_process_id_t peer, int service,
1300 int nbulkiov, int bulklen,
1301 void (*rpc_done)(srpc_client_rpc_t *),
1302 void (*rpc_fini)(srpc_client_rpc_t *), void *priv)
1303{
1304 srpc_client_rpc_t *rpc;
1305
1306 LIBCFS_ALLOC(rpc, offsetof(srpc_client_rpc_t,
1307 crpc_bulk.bk_iovs[nbulkiov]));
1308 if (rpc == NULL)
1309 return NULL;
1310
1311 srpc_init_client_rpc(rpc, peer, service, nbulkiov,
1312 bulklen, rpc_done, rpc_fini, priv);
1313 return rpc;
1314}
1315
1316/* called with rpc->crpc_lock held */
1317void
1318srpc_abort_rpc (srpc_client_rpc_t *rpc, int why)
1319{
1320 LASSERT (why != 0);
1321
1322 if (rpc->crpc_aborted || /* already aborted */
1323 rpc->crpc_closed) /* callback imminent */
1324 return;
1325
1326 CDEBUG (D_NET,
1327 "Aborting RPC: service %d, peer %s, state %s, why %d\n",
1328 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1329 swi_state2str(rpc->crpc_wi.swi_state), why);
1330
1331 rpc->crpc_aborted = 1;
1332 rpc->crpc_status = why;
1333 swi_schedule_workitem(&rpc->crpc_wi);
1334 return;
1335}
1336
1337/* called with rpc->crpc_lock held */
1338void
1339srpc_post_rpc (srpc_client_rpc_t *rpc)
1340{
1341 LASSERT (!rpc->crpc_aborted);
1342 LASSERT (srpc_data.rpc_state == SRPC_STATE_RUNNING);
1343
1344 CDEBUG (D_NET, "Posting RPC: peer %s, service %d, timeout %d\n",
1345 libcfs_id2str(rpc->crpc_dest), rpc->crpc_service,
1346 rpc->crpc_timeout);
1347
1348 srpc_add_client_rpc_timer(rpc);
1349 swi_schedule_workitem(&rpc->crpc_wi);
1350 return;
1351}
1352
1353
1354int
1355srpc_send_reply(struct srpc_server_rpc *rpc)
1356{
1357 srpc_event_t *ev = &rpc->srpc_ev;
1358 struct srpc_msg *msg = &rpc->srpc_replymsg;
1359 struct srpc_buffer *buffer = rpc->srpc_reqstbuf;
1360 struct srpc_service_cd *scd = rpc->srpc_scd;
1361 struct srpc_service *sv = scd->scd_svc;
1362 __u64 rpyid;
1363 int rc;
1364
1365 LASSERT(buffer != NULL);
1366 rpyid = buffer->buf_msg.msg_body.reqst.rpyid;
1367
1368 spin_lock(&scd->scd_lock);
1369
1370 if (!sv->sv_shuttingdown && !srpc_serv_is_framework(sv)) {
1371 /* Repost buffer before replying since test client
1372 * might send me another RPC once it gets the reply */
1373 if (srpc_service_post_buffer(scd, buffer) != 0)
1374 CWARN("Failed to repost %s buffer\n", sv->sv_name);
1375 rpc->srpc_reqstbuf = NULL;
1376 }
1377
1378 spin_unlock(&scd->scd_lock);
1379
1380 ev->ev_fired = 0;
1381 ev->ev_data = rpc;
1382 ev->ev_type = SRPC_REPLY_SENT;
1383
1384 msg->msg_magic = SRPC_MSG_MAGIC;
1385 msg->msg_version = SRPC_MSG_VERSION;
1386 msg->msg_type = srpc_service2reply(sv->sv_id);
1387
1388 rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, rpyid, msg,
1389 sizeof(*msg), LNET_MD_OP_PUT,
1390 rpc->srpc_peer, rpc->srpc_self,
1391 &rpc->srpc_replymdh, ev);
1392 if (rc != 0)
1393 ev->ev_fired = 1; /* no more event expected */
1394 return rc;
1395}
1396
1397/* when in kernel always called with LNET_LOCK() held, and in thread context */
1398void
1399srpc_lnet_ev_handler(lnet_event_t *ev)
1400{
1401 struct srpc_service_cd *scd;
1402 srpc_event_t *rpcev = ev->md.user_ptr;
1403 srpc_client_rpc_t *crpc;
1404 srpc_server_rpc_t *srpc;
1405 srpc_buffer_t *buffer;
1406 srpc_service_t *sv;
1407 srpc_msg_t *msg;
1408 srpc_msg_type_t type;
1409
1410 LASSERT (!in_interrupt());
1411
1412 if (ev->status != 0) {
1413 spin_lock(&srpc_data.rpc_glock);
1414 srpc_data.rpc_counters.errors++;
1415 spin_unlock(&srpc_data.rpc_glock);
1416 }
1417
1418 rpcev->ev_lnet = ev->type;
1419
1420 switch (rpcev->ev_type) {
1421 default:
1422 CERROR("Unknown event: status %d, type %d, lnet %d\n",
1423 rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1424 LBUG ();
1425 case SRPC_REQUEST_SENT:
1426 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1427 spin_lock(&srpc_data.rpc_glock);
1428 srpc_data.rpc_counters.rpcs_sent++;
1429 spin_unlock(&srpc_data.rpc_glock);
1430 }
1431 case SRPC_REPLY_RCVD:
1432 case SRPC_BULK_REQ_RCVD:
1433 crpc = rpcev->ev_data;
1434
1435 if (rpcev != &crpc->crpc_reqstev &&
1436 rpcev != &crpc->crpc_replyev &&
1437 rpcev != &crpc->crpc_bulkev) {
1438 CERROR("rpcev %p, crpc %p, reqstev %p, replyev %p, bulkev %p\n",
1439 rpcev, crpc, &crpc->crpc_reqstev,
1440 &crpc->crpc_replyev, &crpc->crpc_bulkev);
1441 CERROR("Bad event: status %d, type %d, lnet %d\n",
1442 rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1443 LBUG ();
1444 }
1445
1446 spin_lock(&crpc->crpc_lock);
1447
1448 LASSERT(rpcev->ev_fired == 0);
1449 rpcev->ev_fired = 1;
1450 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1451 -EINTR : ev->status;
1452 swi_schedule_workitem(&crpc->crpc_wi);
1453
1454 spin_unlock(&crpc->crpc_lock);
1455 break;
1456
1457 case SRPC_REQUEST_RCVD:
1458 scd = rpcev->ev_data;
1459 sv = scd->scd_svc;
1460
1461 LASSERT(rpcev == &scd->scd_ev);
1462
1463 spin_lock(&scd->scd_lock);
1464
1465 LASSERT (ev->unlinked);
1466 LASSERT (ev->type == LNET_EVENT_PUT ||
1467 ev->type == LNET_EVENT_UNLINK);
1468 LASSERT (ev->type != LNET_EVENT_UNLINK ||
1469 sv->sv_shuttingdown);
1470
1471 buffer = container_of(ev->md.start, srpc_buffer_t, buf_msg);
1472 buffer->buf_peer = ev->initiator;
1473 buffer->buf_self = ev->target.nid;
1474
1475 LASSERT(scd->scd_buf_nposted > 0);
1476 scd->scd_buf_nposted--;
1477
1478 if (sv->sv_shuttingdown) {
1479 /* Leave buffer on scd->scd_buf_nposted since
1480 * srpc_finish_service needs to traverse it. */
1481 spin_unlock(&scd->scd_lock);
1482 break;
1483 }
1484
1485 if (scd->scd_buf_err_stamp != 0 &&
1486 scd->scd_buf_err_stamp < cfs_time_current_sec()) {
1487 /* re-enable adding buffer */
1488 scd->scd_buf_err_stamp = 0;
1489 scd->scd_buf_err = 0;
1490 }
1491
1492 if (scd->scd_buf_err == 0 && /* adding buffer is enabled */
1493 scd->scd_buf_adjust == 0 &&
1494 scd->scd_buf_nposted < scd->scd_buf_low) {
1495 scd->scd_buf_adjust = MAX(scd->scd_buf_total / 2,
1496 SFW_TEST_WI_MIN);
1497 swi_schedule_workitem(&scd->scd_buf_wi);
1498 }
1499
1500 list_del(&buffer->buf_list); /* from scd->scd_buf_posted */
1501 msg = &buffer->buf_msg;
1502 type = srpc_service2request(sv->sv_id);
1503
1504 if (ev->status != 0 || ev->mlength != sizeof(*msg) ||
1505 (msg->msg_type != type &&
1506 msg->msg_type != __swab32(type)) ||
1507 (msg->msg_magic != SRPC_MSG_MAGIC &&
1508 msg->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1509 CERROR ("Dropping RPC (%s) from %s: "
1510 "status %d mlength %d type %u magic %u.\n",
1511 sv->sv_name, libcfs_id2str(ev->initiator),
1512 ev->status, ev->mlength,
1513 msg->msg_type, msg->msg_magic);
1514
1515 /* NB can't call srpc_service_recycle_buffer here since
1516 * it may call LNetM[DE]Attach. The invalid magic tells
1517 * srpc_handle_rpc to drop this RPC */
1518 msg->msg_magic = 0;
1519 }
1520
1521 if (!list_empty(&scd->scd_rpc_free)) {
1522 srpc = list_entry(scd->scd_rpc_free.next,
1523 struct srpc_server_rpc,
1524 srpc_list);
1525 list_del(&srpc->srpc_list);
1526
1527 srpc_init_server_rpc(srpc, scd, buffer);
1528 list_add_tail(&srpc->srpc_list,
1529 &scd->scd_rpc_active);
1530 swi_schedule_workitem(&srpc->srpc_wi);
1531 } else {
1532 list_add_tail(&buffer->buf_list,
1533 &scd->scd_buf_blocked);
1534 }
1535
1536 spin_unlock(&scd->scd_lock);
1537
1538 spin_lock(&srpc_data.rpc_glock);
1539 srpc_data.rpc_counters.rpcs_rcvd++;
1540 spin_unlock(&srpc_data.rpc_glock);
1541 break;
1542
1543 case SRPC_BULK_GET_RPLD:
1544 LASSERT (ev->type == LNET_EVENT_SEND ||
1545 ev->type == LNET_EVENT_REPLY ||
1546 ev->type == LNET_EVENT_UNLINK);
1547
1548 if (!ev->unlinked)
1549 break; /* wait for final event */
1550
1551 case SRPC_BULK_PUT_SENT:
1552 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1553 spin_lock(&srpc_data.rpc_glock);
1554
1555 if (rpcev->ev_type == SRPC_BULK_GET_RPLD)
1556 srpc_data.rpc_counters.bulk_get += ev->mlength;
1557 else
1558 srpc_data.rpc_counters.bulk_put += ev->mlength;
1559
1560 spin_unlock(&srpc_data.rpc_glock);
1561 }
1562 case SRPC_REPLY_SENT:
1563 srpc = rpcev->ev_data;
1564 scd = srpc->srpc_scd;
1565
1566 LASSERT(rpcev == &srpc->srpc_ev);
1567
1568 spin_lock(&scd->scd_lock);
1569
1570 rpcev->ev_fired = 1;
1571 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1572 -EINTR : ev->status;
1573 swi_schedule_workitem(&srpc->srpc_wi);
1574
1575 spin_unlock(&scd->scd_lock);
1576 break;
1577 }
1578}
1579
1580
1581int
1582srpc_startup (void)
1583{
1584 int rc;
1585
1586 memset(&srpc_data, 0, sizeof(struct smoketest_rpc));
1587 spin_lock_init(&srpc_data.rpc_glock);
1588
1589 /* 1 second pause to avoid timestamp reuse */
1590 cfs_pause(cfs_time_seconds(1));
1591 srpc_data.rpc_matchbits = ((__u64) cfs_time_current_sec()) << 48;
1592
1593 srpc_data.rpc_state = SRPC_STATE_NONE;
1594
1595 rc = LNetNIInit(LUSTRE_SRV_LNET_PID);
1596 if (rc < 0) {
1597 CERROR ("LNetNIInit() has failed: %d\n", rc);
1598 return rc;
1599 }
1600
1601 srpc_data.rpc_state = SRPC_STATE_NI_INIT;
1602
1603 LNetInvalidateHandle(&srpc_data.rpc_lnet_eq);
1604 rc = LNetEQAlloc(0, srpc_lnet_ev_handler, &srpc_data.rpc_lnet_eq);
1605 if (rc != 0) {
1606 CERROR("LNetEQAlloc() has failed: %d\n", rc);
1607 goto bail;
1608 }
1609
1610 rc = LNetSetLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1611 LASSERT(rc == 0);
1612 rc = LNetSetLazyPortal(SRPC_REQUEST_PORTAL);
1613 LASSERT(rc == 0);
1614
1615 srpc_data.rpc_state = SRPC_STATE_EQ_INIT;
1616
1617 rc = stt_startup();
1618
1619bail:
1620 if (rc != 0)
1621 srpc_shutdown();
1622 else
1623 srpc_data.rpc_state = SRPC_STATE_RUNNING;
1624
1625 return rc;
1626}
1627
1628void
1629srpc_shutdown (void)
1630{
1631 int i;
1632 int rc;
1633 int state;
1634
1635 state = srpc_data.rpc_state;
1636 srpc_data.rpc_state = SRPC_STATE_STOPPING;
1637
1638 switch (state) {
1639 default:
1640 LBUG ();
1641 case SRPC_STATE_RUNNING:
1642 spin_lock(&srpc_data.rpc_glock);
1643
1644 for (i = 0; i <= SRPC_SERVICE_MAX_ID; i++) {
1645 srpc_service_t *sv = srpc_data.rpc_services[i];
1646
1647 LASSERTF (sv == NULL,
1648 "service not empty: id %d, name %s\n",
1649 i, sv->sv_name);
1650 }
1651
1652 spin_unlock(&srpc_data.rpc_glock);
1653
1654 stt_shutdown();
1655
1656 case SRPC_STATE_EQ_INIT:
1657 rc = LNetClearLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1658 rc = LNetClearLazyPortal(SRPC_REQUEST_PORTAL);
1659 LASSERT (rc == 0);
1660 rc = LNetEQFree(srpc_data.rpc_lnet_eq);
1661 LASSERT (rc == 0); /* the EQ should have no user by now */
1662
1663 case SRPC_STATE_NI_INIT:
1664 LNetNIFini();
1665 }
1666
1667 return;
1668}
Linux kernel - Deliverables
This page took 0.121972 seconds and 5 git commands to generate.